Bayesian parameterisation of a regional photovoltaic model – Application to forecasting. (August 2019)

• Record Type: Journal Article
• Authors: Saint-Drenan, Yves-Marie; Vogt, Stephan; Killinger, Sven; Bright, Jamie M.; Fritz, Rafael; Potthast, Roland
• Is Part Of: Solar energy. Volume 188(2019)
• Page Start: 760
• Page End: 774
• Publication Date: 2019-08
• Subjects: PV system characteristics -- Regional PV power -- Forecast -- Grid integration -- Inverse problem; Solar energy -- Periodicals; Solar engines -- Periodicals; 621.47
• DOI: 10.1016/j.solener.2019.06.053
• Languages: English
• ISSNs: 0038-092X

Abstract

Highlights: A method to train a regional PV model from the total power production is proposed. A regularization of the parameter estimation increases the model accuracy. A validation of the approach against TSO data shows promising results. Abstract: Estimating and forecasting photovoltaic (PV) power generation in regions—e.g. the area controlled by the transmission system operator (TSO)—is a requirement for the operation of the electricity supply system. An accurate calculation of this quantity requires detailed information of the installed PV systems within the considered region; however, this information is not publicly available making forecasting difficult. Therefore, approximating the undefined PV systems information for use in a PV power model (parameterization) is of critical interest. In this paper, we propose a methodological approach for parameterization using time series of aggregated PV power generation. A Bayesian approach is used to overcome the significant number of unknown parameters in the problem. It regularizes the linear system by imposing constraints on deviations from an initial-guess and covariance matrices; the initial guess uses available statistical distributions of PV system metadata. The performance of the proposed forecasting approach is evaluated using estimates of the regional PV power generation from three TSOs and meteorological data from the IFS forecast model (ECMWF). The proposed forecasting approach without the Bayesian parameterization has RMSE of 3.90%, 4.25% and 4.64%, respectively; including the Bayesian approach gives RMSE of 3.82%, 4.23% and 4.51%. For comparison, we also deployed a multiple linear regression which gave RMSE of 3.89%, 4.12% and 4.54%; however, there are considerable downsides to such an approach. Our approach is competitive with TSO forecasting systems despite using far fewer input data and simpler implementation of NWP prediction. This is particularly promising as there are many avenues for future development.